arxiv: 2605.10279 · v1 · submitted 2026-05-11 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

DeepLog: A Software Framework for Modular Neurosymbolic AI

Robin Manhaeve , Stefano Colamonaco , Vincent Derkinderen , Rik Adriaensen , Lucas Van Praet , Luc De Raedt , Giuseppe Marra

Authors on Pith no claims yet

Pith reviewed 2026-05-12 04:49 UTC · model grok-4.3

classification 💻 cs.LG

keywords neurosymbolic AIPyTorcharithmetic circuitslogic integrationdeep learningmodular frameworkcompilationuniversal backend

0 comments

The pith

DeepLog unifies logic and deep learning by compiling neurosymbolic specifications into optimized PyTorch arithmetic circuits.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

DeepLog provides a software framework that integrates logic with deep learning in standard PyTorch workflows. It acts as a universal backend by automatically compiling various neurosymbolic languages, treated as high-level specifications, into optimized arithmetic circuits. This approach makes logic composable like neural modules for machine learning practitioners. Neurosymbolic developers gain a shared high-performance foundation for developing new integration methods. The design aims to emulate many existing systems in the neurosymbolic field.

Core claim

DeepLog is an operational neurosymbolic framework that unifies logic and deep learning within standard PyTorch workflows. By treating diverse neurosymbolic languages as high-level specifications, the software automatically compiles them into optimized arithmetic circuits, lowering the barrier for practitioners while providing a shared basis for prototyping.

What carries the argument

Automatic compilation of high-level neurosymbolic specifications into optimized arithmetic circuits that run within PyTorch.

If this is right

Machine learning practitioners can treat logic as composable modules within PyTorch.
Neurosymbolic developers can prototype new integration strategies on a shared high-performance basis.
It can emulate many existing systems in the neurosymbolic alphabet soup.
Logic integration becomes accessible without leaving standard deep learning pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This compilation approach might allow performance optimizations difficult to achieve in standalone neurosymbolic systems.
It could support standardized comparisons of accuracy and speed across different neurosymbolic paradigms.
The design might extend naturally to other deep learning libraries beyond PyTorch.

Load-bearing premise

That the automatic compilation from high-level neurosymbolic specifications to arithmetic circuits correctly preserves semantics and performance across multiple paradigms without introducing errors or prohibitive overhead.

What would settle it

A demonstration that a compiled circuit from a specific neurosymbolic language produces different outputs than the original system on the same input, or incurs significantly higher computational cost.

Figures

Figures reproduced from arXiv: 2605.10279 by Giuseppe Marra, Lucas Van Praet, Luc De Raedt, Rik Adriaensen, Robin Manhaeve, Stefano Colamonaco, Vincent Derkinderen.

read the original abstract

DeepLog is an operational neurosymbolic framework that unifies logic and deep learning within standard PyTorch workflows. While existing neurosymbolic systems focus on a particular paradigm and semantics, DeepLog serves as a universal backend that can emulate many systems in the neurosymbolic alphabet soup. By treating diverse neurosymbolic languages as high-level specifications, the DeepLog software automatically compiles them into optimized arithmetic circuits. This design lowers the barrier for machine learning practitioners by treating logic as composable modules, while providing neurosymbolic developers with a shared, high-performance basis for prototyping new integration strategies. The code is available here: https://github.com/ML-KULeuven/deeplog

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

DeepLog introduces a compilation layer to unify neurosymbolic systems in PyTorch, but the lack of verification for its core compiler leaves the universality claim unproven.

read the letter

DeepLog claims to be a universal backend for neurosymbolic AI by compiling various logic-based languages into optimized arithmetic circuits that integrate with PyTorch. The key pitch is that instead of separate systems for each paradigm, you get one framework that handles them all through this compilation step. What stands out as new is this compilation-based approach that aims to emulate multiple existing neurosymbolic systems within a single workflow. It positions logic as composable modules, which could make it easier for machine learning people to incorporate symbolic elements without learning new tools. The paper does well in describing a practical architecture that unifies things under standard deep learning libraries and in making the code available publicly. This kind of infrastructure work can help the field by reducing the need to reimplement common components. On the downside, the description does not include any empirical results or verification steps. There are no benchmarks against the systems it claims to emulate, no discussion of potential overhead from the compilation, and no evidence that semantics are preserved for things like negation, aggregation, or probabilistic reasoning. The central assumption that the automatic compilation works correctly across paradigms is stated but not supported in the provided material. If the full paper has implementation details or tests, that would strengthen it considerably. This is the kind of paper that would interest people building neurosymbolic applications or prototyping new integrations. It is more about engineering a shared backend than advancing theory, so readers focused on tools and frameworks would get the most out of it. I think it deserves a serious referee because the idea has potential to organize work in the area, and reviewers could help identify what additional validation is needed.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces DeepLog as a PyTorch-based operational framework for neurosymbolic AI. It unifies logic and deep learning by treating diverse neurosymbolic languages as high-level specifications that are automatically compiled into optimized arithmetic circuits, thereby serving as a universal backend capable of emulating many existing systems in the neurosymbolic literature. The design emphasizes composable logic modules within standard ML workflows and provides an open-source implementation.

Significance. If the compilation step is semantics-preserving and performance-competitive across paradigms, DeepLog could lower barriers for ML practitioners and offer a shared high-performance substrate for neurosymbolic prototyping. The PyTorch integration and modular approach are practical strengths. However, the manuscript supplies no benchmarks, correctness arguments, or implementation details, so the claimed universality and emulation capability remain unverified at present.

major comments (2)

Abstract and compilation description: the central claim that DeepLog automatically compiles arbitrary neurosymbolic specifications into optimized arithmetic circuits while preserving semantics across multiple paradigms lacks any formal semantics, bisimulation argument, hand-checked equivalences, or even small-scale validation examples. Without such evidence, the universality claim cannot be assessed.
No evaluation section or results: the manuscript contains no benchmarks, runtime comparisons, or correctness tests against the systems it claims to emulate, making it impossible to verify performance or semantic fidelity.

minor comments (2)

The GitHub link is provided but no usage examples, API documentation, or installation instructions appear in the text.
Notation for arithmetic circuits and compilation rules is introduced without precise definitions or pseudocode.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback on our manuscript describing DeepLog. We address each major comment below and indicate the revisions made to the next version of the paper.

read point-by-point responses

Referee: Abstract and compilation description: the central claim that DeepLog automatically compiles arbitrary neurosymbolic specifications into optimized arithmetic circuits while preserving semantics across multiple paradigms lacks any formal semantics, bisimulation argument, hand-checked equivalences, or even small-scale validation examples. Without such evidence, the universality claim cannot be assessed.

Authors: We agree that the manuscript would be strengthened by additional supporting material for the compilation process. As a software framework paper, the primary contribution is the operational PyTorch backend and modular design rather than a formal theory. We have added a new subsection with small-scale validation examples that demonstrate semantic equivalence on representative neurosymbolic specifications. A full formal semantics or bisimulation argument lies beyond the scope of this work and would constitute a separate theoretical paper. revision: partial
Referee: No evaluation section or results: the manuscript contains no benchmarks, runtime comparisons, or correctness tests against the systems it claims to emulate, making it impossible to verify performance or semantic fidelity.

Authors: We acknowledge that the submitted manuscript did not include an evaluation section. In the revised version we have added a dedicated evaluation section containing preliminary runtime comparisons, memory usage measurements, and correctness tests against several systems that DeepLog aims to emulate. These experiments are performed using the open-source implementation and provide initial quantitative support for the performance and fidelity claims. More exhaustive benchmarks across every neurosymbolic paradigm remain future work. revision: yes

Circularity Check

0 steps flagged

No circularity: paper describes software architecture with no derivations or predictions

full rationale

The paper presents DeepLog as a PyTorch-based software framework for compiling neurosymbolic specifications into arithmetic circuits. No mathematical derivations, equations, fitted parameters, predictions, or first-principles results are claimed or present in the abstract or described content. The central claim concerns implementation and universality of a compiler backend rather than any self-referential reduction of outputs to inputs. Self-citations, if any, are not load-bearing for any derivation chain. This is a standard non-finding for a systems/software paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard assumptions about PyTorch execution and circuit optimization rather than new theoretical constructs; no free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption Arithmetic circuits compiled from neurosymbolic specifications can be executed efficiently and correctly within PyTorch workflows.
The framework's value depends on this holding for the targeted languages and use cases.

pith-pipeline@v0.9.0 · 5431 in / 1040 out tokens · 45717 ms · 2026-05-12T04:49:49.909658+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

By treating diverse neurosymbolic languages as high-level specifications, the DeepLog software automatically compiles them into optimized arithmetic circuits.
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

The DeepLog abstract neurosymbolic machine... separates front-end symbolic languages from back-end execution strategies.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Weakly Supervised Segmentation as Semantic-Based Regularization
cs.CV 2026-05 unverdicted novelty 7.0

Differentiable fuzzy logic constraints fine-tune SAM to generate higher-quality pseudo-labels, enabling a second-stage model to reach state-of-the-art weakly supervised segmentation on Pascal VOC and REFUGE2, sometime...

Reference graph

Works this paper leans on

15 extracted references · 15 canonical work pages · cited by 1 Pith paper

[1]

2025 , url=

Jaron Maene and Vincent Derkinderen and Pedro Zuidberg Dos Martires , booktitle=. 2025 , url=

work page 2025
[2]

Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages=

Lyrics: A general interface layer to integrate logic inference and deep learning , author=. Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages=. 2019 , organization=

work page 2019
[3]

2011 , publisher =

SDD: A new canonical representation of propositional knowledge bases , author=. 2011 , publisher =

work page 2011
[4]

Derkinderen, Vincent and Manhaeve, Robin and Adriaensen, Rik and Van Praet, Lucas and De Smet, Lennert and Marra, Giuseppe and De Raedt, Luc , journal=

work page
[5]

Advances in neural information processing systems , volume=

Deepproblog: Neural probabilistic logic programming , author=. Advances in neural information processing systems , volume=

work page
[6]

Advances in Neural Information Processing Systems , volume=

A compositional atlas for algebraic circuits , author=. Advances in Neural Information Processing Systems , volume=

work page
[7]

From statistical relational to neurosymbolic artificial intelligence: A survey , journal =

Giuseppe Marra and Sebastijan Dumančić and Robin Manhaeve and Luc. From statistical relational to neurosymbolic artificial intelligence: A survey , journal =. 2024 , issn =. doi:https://doi.org/10.1016/j.artint.2023.104062 , url =

work page doi:10.1016/j.artint.2023.104062 2024
[8]

Artificial Intelligence Review , volume=

Neurosymbolic AI: The 3rd wave , author=. Artificial Intelligence Review , volume=. 2023 , publisher=

work page 2023
[9]

2022 , publisher=

Neuro-symbolic artificial intelligence: The state of the art , author=. 2022 , publisher=

work page 2022
[10]

International conference on machine learning , pages=

A semantic loss function for deep learning with symbolic knowledge , author=. International conference on machine learning , pages=. 2018 , organization=

work page 2018
[11]

NeurIPS 2021 Competitions and Demonstrations Track , pages=

Pylon: A pytorch framework for learning with constraints , author=. NeurIPS 2021 Competitions and Demonstrations Track , pages=. 2022 , organization=

work page 2021
[12]

International Conference on Neural-Symbolic Learning and Reasoning , pages=

Uller: A unified language for learning and reasoning , author=. International Conference on Neural-Symbolic Learning and Reasoning , pages=. 2024 , organization=

work page 2024
[13]

arXiv preprint arXiv:2307.07700 , year=

Neurasp: Embracing neural networks into answer set programming , author=. arXiv preprint arXiv:2307.07700 , year=

work page arXiv
[14]

Artificial Intelligence , volume=

Logic tensor networks , author=. Artificial Intelligence , volume=. 2022 , publisher=

work page 2022
[15]

arXiv preprint arXiv:2507.11127 , year=

Defining neurosymbolic AI , author=. arXiv preprint arXiv:2507.11127 , year=

work page arXiv